…s.ts

McpToolContext was defined in server.ts, which imported TOOL_HANDLERS from
tools/index.ts (the barrel). Every tool module imported McpToolContext back
from server.ts, creating a 37-file circular dependency flagged by codegraph
cycles in two consecutive architectural audits.

Fix: extract McpToolContext and McpToolHandler into src/mcp/types.ts, which
only depends on db/index.js (outside the MCP subtree). server.ts and all 35
tool modules now import from types.ts instead of server.ts, eliminating the
cycle. server.ts re-exports McpToolContext for backward compatibility.

Replace `any` return types with `typeof Database` from the installed
@types/better-sqlite3 package in src/db/better-sqlite3.ts and
src/mcp/types.ts, completing the migration away from hand-rolled
better-sqlite3 type declarations.

Closes #1622

…-native edges

Two-part fix for the confidence regression introduced when the ts-native
resolution pass added 12,776 cross-module edges at 0.3 confidence:

1. Exclude sink edges (confidence=0.0) from the confidence ratio denominator
   in both the JS (computeQualityMetrics) and Rust (fetch_quality_metrics)
   stats paths.  Sink edges flag unresolvable dynamic calls (eval/computed-key)
   and are not resolution attempts — counting them against resolution quality
   was incorrect.  The FP ratio still uses the full edge count.

2. Lift the minimum confidence for ts-native resolved edges from 0.3 → 0.5
   (TS_NATIVE_CONFIDENCE_FLOOR).  The proximity heuristic returns 0.3 for
   cross-module calls where no import-path evidence is available, but the
   native and WASM engines both perform actual name-based symbol lookup —
   stronger evidence than pure file-proximity.  0.5 (same-parent-directory
   level) is a conservative but correct floor.  Sink edges (confidence=0.0)
   are explicitly excluded from the lift.

The floor is applied: in-memory to allEdgeRows before batchInsertEdges (WASM
and fallback paths); via SQL UPDATE in applyEdgeTechniquesAfterNativeInsert
(native bulk-insert path); via SQL UPDATE in
backfillEdgeTechniquesAfterNativeOrchestrator (native orchestrator path).

Closes #1623

…ison table

…nstructors

Rust struct constructors (e.g. Import.new, FileSymbols.new) always have
high call counts by design. Add isStructuralFalsePositive() to exclude
any method named "new" in a .rs file from the false-positive warning
list — pattern-based so no specific symbol names are hardcoded.

Impact: 2 functions changed, 6 affected

… WASM backfill

The native Rust engine uses the `ignore` crate with git_ignore(true) so it
respects .gitignore and never processes gitignored files. The JS collectFiles
walker has no gitignore awareness, so gitignored files such as
crates/codegraph-core/index.js and index.d.ts (NAPI-RS generated) appeared in
the expected-file set but not in the DB after a native build, triggering the
spurious "Native orchestrator dropped N file(s) — likely a Rust extractor bug"
WARN and an unnecessary WASM backfill on every fresh build.

Fix: add queryGitIgnoredFiles() which runs `git check-ignore --stdin` on the
candidate file list. detectDroppedLanguageGap() now filters the expected set
through this check so gitignored files are never treated as native extractor
failures. Falls back gracefully when git is unavailable or the directory is not
a git repository.

Closes #1626

Impact: 2 functions changed, 6 affected

…repos

The gitignore-based filter in queryGitIgnoredFiles already excludes the
NAPI-RS binding artifacts (crates/codegraph-core/index.js / index.d.ts)
from gap detection. Adding 'crates' to IGNORE_DIRS is therefore redundant
for this repo and harmful for any other repo that has a legitimately named
crates/ directory with indexable source files.